US20120029662A1

US20120029662A1 - Advanced process control system and method utilizing virtual metrology with reliance index

Info

Publication number: US20120029662A1
Application number: US13/193,607
Authority: US
Inventors: Fan-Tien Cheng; Chi-An Kao; Wei-Ming Wu
Original assignee: National Cheng Kung University NCKU; Foresight Technology Co Ltd
Current assignee: National Cheng Kung University NCKU; Foresight Technology Co Ltd
Priority date: 2010-08-02
Filing date: 2011-07-29
Publication date: 2012-02-02
Also published as: KR101335896B1; JP2012033170A; JP5292602B2; CN102375413A; KR20120012766A; TWI427722B; US8688256B2; TW201212140A; CN102375413B

Abstract

An advanced process control (APC) system, an APC method, and a computer program product, which, when executed, performs an APC method are provided for incorporating virtual metrology (VM) into APC. The present inventions uses a reliance index (RI) and a global similarity index (GSI) to adjust at least one controller gain of a run-to-run (R2R) controller when the VM value of a workpiece is adopted to replace the actual measurement value of the workpiece. The RI is used for gauging the reliability of the VM value, and the GSI is used for assessing the degree of similarity between the set of process data for generating the VM value and all the sets of historical process data used for building the conjecturing model.

Description

RELATED APPLICATIONS

The present application is based on, and claims priority from, U.S. provisional Application Ser. No. 61/369,761, filed Aug. 2, 2010, the disclosure of which is hereby incorporated by reference herein in its entirety.

BACKGROUND

1. Field of Invention
The present invention relates to an advanced process control (APC) system and an APC method. More particularly, the present invention relates to an APC system and an APC method utilizing virtual metrology (VM) with a reliance index (RI).
2. Description of Related Art
Run-to-run (R2R) advanced process control (APC) is widely applied to semiconductor and TFT-LCD factories for improving process capability. As defined in SEMI E133 specification, a R2R control is the technique of modifying recipe parameters or the selection of control parameters between runs to improve processing performance. A (process) run can be a batch, a lot, or an individual workpiece, wherein the R2R APC becomes a lot-to-lot (L2L) APC when a run is a lot, and the R2R APC becomes a workpiece-to-workpiece (W2W) APC when a run is a workpiece. A workpiece may represent a wafer for the semiconductor industry or a glass for the TFT-LCD industry. The L2L APC is now widely implemented for dealing with advanced technologies. When a L2L control is applied, only one single workpiece in the lot is required to be measured for feedback and feedforward control purposes. However, as the device dimension shrinks further, tighter process control is needed. In this case, the L2L control may not be accurate enough and therefore a W2W control becomes essential for critical stages. As a result, each workpiece in the lot should be measured. To fulfill this requirement, large amounts of metrology tools will be required and production cycle time will also be increased significantly. Furthermore, metrology delays, which are inevitable as real measurements are performed, will not only cause complicated control problems but also degrade the APC performance.
To resolve the problem mentioned above, virtual metrology (VM) was proposed. Virtual metrology is a technology using a conjecture model to predict metrology variables using information about the state of the process for every workpiece. If the VM conjecture model is fresh and accurate enough, it can generate a VM value within seconds after collecting the complete tool process data of a workpiece. Therefore, this VM value can be applied for real-time W2W control.
Referring to FIG. 1, FIG. 1 is a schematic block diagram showing a conventional model of EWMA (Exponentially Weighted Moving Average) R2R control disclosed by the paper entitled “Performance Analysis of EWMA Controllers Subject to Metrology Delay”, M.-F. Wu, C.-H. Lin, D. S.-H. Wong, S.-S. Jang, and S.-T. Tseng, published in IEEE Transactions on Semiconductor Manufacturing, vol. 21, no. 3, pp. 413-425, August 2008, which is incorporated herein by reference. Let us consider a process model with linear input and output relationship:
y _k=β₀+β₁ u _k+η_k (1)
where y_kis the plant output; u_kthe control action taken for process run k; β₀the initial bias of process; β₁the process gain; and η₅the disturbance model input.
Given a process predictive model Au_k, where A is a gain parameter (e.g., removal rate for chemical mechanical polishing (CMP)) estimated for the system, and its initial values can be obtained from the actual tool/recipe performance.
Using an EWMA (Exponentially Weighted Moving Average) filter, the model offset or disturbance of the (k+1)^thprocess run is estimated to be
{tilde over (η)}_k+1=α(y _k −Au _k)+(1−α){tilde over (η)}_k (2)
where α is an EWMA coefficient ranged between 0 and 1.
Control action of (k+1)^thprocess run is
$\begin{matrix} u_{k + 1} = \frac{Tgt - {\tilde{η}}_{k + 1}}{A} & (3) \end{matrix}$
where Tgt represents the target value.
Referring to FIG. 2, FIG. 2 is a schematic block diagram showing a conventional W2W control scheme utilizing virtual metrology (VM), wherein y_zis the z^thprocess run of actual measurement value of the sampling product (workpiece) measured by a metrology tool 20; ŷ_kis the k^thprocess run of VM data; and X_kis the k^thprocess run of process data of a process tool 10. In the paper entitled “On the Quality of Virtual Metrology Data for Use in the feedback Process Control”, A. A. Khan, J. R. Moyne, and D. M. Tilbury, published in Proc. AEC/APC Symposium XIX—North America, Palm Springs, Calif. USA, September 2007; the paper entitled “An Approach for Factory-Wide Control Utilizing Virtual Metrology”, A. A. Khan, J. R. Moyne, and D. M. Tilbury, published in IEEE Transactions on Semiconductor Manufacturing, vol. 20, no. 4, pp. 364-375, November 2007; and the paper entitled “Virtual Metrology and Feedback Control for Semiconductor Manufacturing Process Using Recursive Partial Least Squares”, A. A. Khan, J. R. Moyne, and D. M. Tilbury, published in Journal of Process Control, vol. 18, pp. 961-974, 2008, which are incorporated herein by reference, Khan et al. proposed a W2W control scheme utilizing VM. Khan et al. proposed to modify the above equation (2) for a R2R controller 40 as follows:
When y_kis measured by the actual metrology tool 20, it becomes y_z, an EWMA coefficient α, is used in
{tilde over (η)}_k+1=α₁(y _z −Au _k)+(1−α₁){tilde over (η)}_k (4)
When y_kis conjectured or predicted by a VM module 30, it becomes ŷ_k, i.e. a VM value ŷ_kand an EWMA coefficient α₂is used in
{tilde over (η)}_k+1=α₂(ŷ _k −Au _k)+(1−α₂){tilde over (η)}_k(5)
Khan et al. pointed out that α₁>α₂(usually, depending on the relative quality of virtual metrology data). Now, the controller-gain problem of applying VM is focused on how to set α₂, wherein the rule of thumb is that α₂should depend on the quality or reliability of VM and α₂<α₁. Khan et al. proposed two VM quality metrics to consider incorporating VM quality into the controller gain of a R2R controller 40:

- 1. Prediction error at metrology runs: Error=y−ŷ (6)
- 2. If y and ŷ are zero-mean Gaussian deviations from targets, then Min mean-square-error (MSE) estimator of y based on ŷ is

$\begin{matrix} y_{mmse} = ρ \frac{σ_{y}}{σ_{\hat{y}}} & (7) \end{matrix}$
where the correlation coefficient
$\begin{matrix} ρ = \frac{cov [y, \hat{y}]}{σ_{y} σ_{\hat{y}}} & (8) \end{matrix}$
and σ_yand σ_ŷare standard deviations of y and ŷ, respectively.
Nevertheless, both metrics proposed above have the following disadvantages:

- 1. Both equations (6) & (7) need actual metrology data “y”; however, if actual metrology data (measurement values) (y) are available, then no virtual metrology value (ŷ) are needed at all.
- 2. The value of y_mmsemay not be normalized to be between 0˜1 because y_mmsemay be positive or negative due top ρ.

As a result, it may not be easy to combine the data quality metrics as in equations (6) and (7) into the R2R model. Hence, there is a need to develop an APC system and an APC method utilizing VM with a reliance index (RI) and a global similarity index (GSI) for effectively considering the data quality of VM into the R2R controller.

SUMMARY

An object of the present invention is to provide an APC system and an APC method for effectively considering the data quality of VM into a R2R controller, thereby overcoming the problems of inability to consider the reliance level in the VM feedback loop of R2R control and metrology delays as well as upgrading the APC performance.
According to an aspect of the present invention, an APC system includes a process tool, a metrology tool, a virtual metrology (VM) module, a reliance index (RI) module and a run-to-run (R2R) controller. The process tool is operated for processing a plurality of historical workpieces in accordance with a plurality of sets of historical process data, and performing a plurality of process runs on a plurality of workpieces in accordance with a plurality of sets of process data. The metrology tool is operated for measuring the historical workpieces and a plurality of sampling workpieces selected from the workpieces, thereby providing a plurality of historical measurement data of the historical workpieces and a plurality of actual measurement values of the sampling workpieces which have been processed in the process runs. The virtual metrology module is used for providing a plurality of virtual metrology values of the process runs by inputting the sets of process data into a conjecture model, wherein the conjecture model is built in accordance with a conjecture algorithm by using the sets of historical process data and the historical measurement values, wherein the historical measurement values are the measurement values of the historical workpieces which are manufactured in accordance with the sets of historical process data, respectively. The RI module is used for generating respective reliance indexes (RI) of the process runs, wherein each of the reliance indexes (RI) corresponding to the process run is generated by calculating the overlap area between the statistical distribution of the virtual metrology value of the workpiece and the statistical distribution of a reference prediction value of the workpiece, wherein the reference prediction value of the process run is generated by inputting the set of process data into a reference model, wherein the reference model is built in accordance with a reference algorithm by using the sets of historical process data and their corresponding historical measurement values, and the conjecture algorithm is different from the reference algorithm, and the reliance index is higher when the overlap area is larger, representing that the reliance level of the virtual metrology value corresponding to the reliance index is higher. The R2R controller is operated for controlling the process tool to perform the process runs in accordance with the following relationships:
u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z)
u _k+1 =g(G _2,1 ,G _2,2 , . . . ,G _2,i ,ŷ _k)
G _2,i =f(RI _k)×G _1,i
where G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller, if RI_k<RI_T;
f(RI _k)=RI _k, if RI_k≧RI_Tand k≦C;
f(RI _k)=1−RI _k, if RI_k≧RI_Tand k>C;
wherein y_zrepresents the actual measurement value of the sampling workpiece which has been processed in the z^thprocess run; u_z+1represents the control action of the (z+1)^thprocess run when y_zis adopted; G_1,i, represents the controller gain used in the R2R controller when y_zis adopted, wherein i represents the number of the controller gains used in the R2R controller; ŷ_krepresents the virtual metrology value of the workpiece which has been processed in the k^thprocess run; u_k+1represents the control action of the (k+1)^thprocess run when ŷ_kis adopted; G_2,i,represents the controller gain used in the R2R controller when ŷ_kis adopted; RI_krepresents the reliance index (RI) of the k^thprocess run; RI_Trepresents the RI threshold value based on a maximal tolerable error limit defined by the errors of the virtual metrology values obtained from the conjecture model; and C stands for a predetermined number of process runs.
In one embodiment, the APC system further includes a global similarity index (GSI) module for generating respective global similarity indexes (GSI) of the process runs by inputting the sets of process data into a statistical distance model, wherein the statistical distance model is built in accordance with a statistical distance algorithm by using the sets of historical process data, wherein G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller, if GSI_k>GSI_T, where GSI_krepresents the global similarity index (GSI) of the kth process run; GSI_Trepresents a GSI threshold value defined by two to three times of the maximal global similarity indexes of the sets of historical process data.
According to another aspect of the present invention, in an APC method, a step is performed for obtaining a plurality of sets of historical process data used by a process tool for processing a plurality of historical workpieces. Another step is performed for obtaining a plurality of historical measurement data of the historical workpieces measured by a metrology tool. Another step is performed for establishing a conjecture model in accordance with a conjecture algorithm by using the sets of historical process data and the historical measurement values, wherein the historical measurement values are the measurement values of the historical workpieces which are manufactured in accordance with the sets of historical process data, respectively; and establishing a reference model in accordance with a reference algorithm by using the sets of historical process data and their corresponding historical measurement values, wherein the conjecture algorithm is different from the reference algorithm. Another step is performed for enabling a run-to-run (R2R) controller to control the process tool to perform the process runs in accordance with the aforementioned relationships.
In one embodiment, the APC method further includes establishing a statistical distance model in accordance with a statistical distance algorithm by using the sets of historical process data; and enabling the R2R controller to control the process tool to perform the process runs in accordance with the relationship: G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller, if GSI_k>GSI_T, where GSI_krepresents the global similarity index (GSI) of the kth process run; GSI_Trepresents a GSI threshold value defined by two to three times of the maximal global similarity indexes of the sets of historical process data.
According to another aspect of the present invention, a computer program product is provided and performs the aforementioned APC method when executed.
Hence, with the application of the embodiments of the present invention, the data quality of VM can be effectively considered into the R2R model, thereby overcoming the problems of inability to consider the reliance level in the VM feedback loop of R2R control and metrology delays as well as upgrading the APC performance.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other features, aspects, and advantages of the present invention will become better understood with regard to the following description, appended claims, and accompanying drawings where:

FIG. 1 is a schematic block diagram showing a conventional model of EWMA R2R control;

FIG. 2 is a schematic block diagram showing a conventional W2W control;

FIG. 3A is a schematic diagram showing a W2W APC system according to an embodiment of the present invention;

FIG. 3B is a schematic diagram showing an EWMA controller according to an embodiment of the present invention;

FIG. 4A is a schematic diagram for defining the reliance index (RI) used in an embodiment of the present invention;

FIG. 4B is a schematic diagram for defining the RI threshold (RI_T) according to the embodiment of the present invention;

FIG. 5 is a schematic flow chart showing a W2W APC method according to an embodiment of the present invention;

FIG. 6A to FIG. 6E show the simulation results of five cases for the first 400 workpieces;

FIG. 7 shows the simulation results of five cases for the 45^th-55^thworkpieces; and

FIG. 8 shows the simulation results of five cases for the 344^th-354^thworkpieces.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

Reference will now be made in detail to the embodiments of the present invention, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers are used in the drawings and the description to refer to the same or like parts.
Referring to FIG. 3A, FIG. 3A is a schematic diagram showing an APC system according to an embodiment of the present invention. The APC system of this embodiment includes a process tool 100, a metrology tool 110, a virtual metrology (VM) module 120, a reliance index (RI) module 122, a global similarity index (GSI) module 124 and a run-to-run (R2R) controller 130. The process tool 100 is operated for processing a plurality of historical workpieces in accordance with a plurality of sets of historical process data, and also can be operated for performing a plurality of process runs on a plurality of workpieces in accordance with a plurality of sets of process data. A process run is the unit controlled by the R2R controller 130, wherein when a process run is a lot, the R2R controller 130 is a L2L controller which controls the process tool 100 lot by lot; when a process run is a workpiece, the R2R controller 130 is a W2W controller which controls the process tool 100 workpiece by workpiece. In general, a lot contains a plurality of workpieces, for example, 25 workpieces, meaning that the L2L controller controls one process run (lot) to process 25 workpieces with one set of process data. The metrology tool 110 is operated for measuring the historical workpieces and a plurality of sampling workpieces selected from the workpieces, thereby providing a plurality of historical measurement data of the historical workpieces and a plurality of actual measurement values of the sampling workpieces which have been processed in the process runs.
For the VM module 120, the RI module 122 and the GSI module 124, a conjecture model, a reference model and a statistical distance model are required to be built. The conjecture model is built in accordance with a conjecture algorithm by using the sets of historical process data and the historical measurement values, wherein the historical measurement values are the measurement values of the historical workpieces which are manufactured in accordance with the sets of historical process data, respectively; the reference model is built in accordance with a reference algorithm by using the sets of historical process data and their corresponding historical measurement values; and the statistical distance model is built in accordance with a statistical distance algorithm by using the sets of historical process data. The conjecture algorithm and the reference algorithm can be a multi-regression (MR) algorithm, a support-vector-regression (SVR) algorithm, a neural-networks (NN) algorithm, a partial-least-squares regression (PLSR) algorithm, or a Gaussian-process-regression (GPR) algorithm. The statistical distance model can be a Mahalanobis-distance algorithm or an Euclidean-distance algorithm. The aforementioned algorithms are merely stated as examples, and certainly other algorithms may be applicable to the present invention. The RI and GSI used in the embodiment of the present invention can be referred to U.S. Pat. No. 7,593,912 entitled “Method for evaluating reliance level of a virtual metrology system in product manufacturing”, which is incorporated herein by reference. The RI, GSI and VM models used in the embodiment of the present invention can be referred to U.S. Pat. No. 7,603,328 entitled “Dual-phase virtual metrology method”; and US Patent Publication No. 20090292386 entitled “System and Method for Automatic Virtual Metrology”, which are incorporated herein by reference. It is noted that U.S. Pat. No. 7,593,912, No. 7,603,328 and US Patent Publication No. 20090292386 all have the same assignee as this application.
The VM module 120 is used for providing a plurality of virtual metrology (VM) values of the process runs by inputting the sets of process data into the conjecture model. The RI module 122 is used for generating respective reliance indexes (RI) of the process runs, wherein each of the reliance indexes (RI) corresponding to the process run is generated by calculating the overlap area between the statistical distribution of the virtual metrology value of the workpiece and the statistical distribution of a reference prediction value of the workpiece, wherein the reference prediction value of the process run is generated by inputting the set of process data into the reference model. The RI module 122 mainly uses another algorithm (reference algorithm) to gauge the reliance level of the conjecture algorithm, and thus the conjecture algorithm and the reference algorithm can be any algorithms as long as the conjecture algorithm is different from the reference algorithm. The reliance index is higher when the aforementioned overlap area is larger, representing that the reliance level of the virtual metrology value corresponding to the reliance index is higher. In this embodiment, a RI threshold value (RI_T) is based on a maximal tolerable error limit defined by the errors of the virtual metrology values obtained from the conjecture model. The GSI module 124 is used for generating respective global similarity indexes (GSI) of the process runs by inputting the sets of process data into the statistical distance model. The GSI assesses the degree of similarity between any set of process data and the model set of process data (for example, the historical process data). In this embodiment, a GSI threshold value (GSI_T) is defined by two to three times of the maximal global similarity indexes of the sets of historical process data.
Hereinafter, the R2R controller 130 is exemplified as an EWMA controller for explanation, but the R2R controller 130 also can be a moving-average (MA) controller, double-EWMA controller (d-EWMA) or a proportional-integral-derivative (PID) controller.
Referring to FIG. 3B, FIG. 3B is a schematic diagram showing an EWMA controller according to an embodiment of the present invention. Embodiments of the present invention are featured in overcoming the controller-gain problem of applying VM regarding how to set the EWMA coefficient α₂in equation (5). The rule of thumb is that α₂should depend on the quality or reliability of VM values and α₂<α₁. The embodiments of the present invention use a reliance index (RI) and a global similarity index (GSI) to gauge the quality or reliability of VM values. Because the value of RI is a good VM reliability evaluation index and 0<RI<1, higher RI means better VM reliability, and then the EWMA coefficient α₂can be naturally set as follows:
α₂ =RI×α ₁ (9)
wherein the EWMA coefficient α₁is the same as the α of equation (2).
Equation (9) will be applied when the R2R controller 130 needs relatively high gain. The situations that need high controller gain are: y_kis apart from the target value or the production process is relatively unstable. On the contrary, if y_kis near the target or the production process is relatively stable, then the controller gain should be small. For generating a small controller gain, the EWMA coefficient α₂also can be set as follows:
α₂=(1−RI)×α₁ (10)
Equations (9) and (10) are valid only when RI is good enough; in other words, RI should be greater than RI_T. If RI<RI_T, this VM value cannot be adopted for tuning the R2R controller gain. Further, due to the fact that the GSI is designed to help the RI gauge the reliance level of VM, when GSI>GSI_T, its corresponding VM value cannot be adopted, either. In conclusion, if RI<RI_Tor GSI>GSI_T, then α₂is set to be zero (0).
The issue of the R2R controller-gain management in real-production environment whenever a modification is performed on the process tool 100 is considered as follows. In general, the production process of the first lot (just after a modification is performed) is relatively unstable; therefore, the controller gain should be relatively high. After finishing the production of the first lot, the production process will become comparatively stable. In other words, the rest of the lots should have small controller gains.
In summary, α₂can be set as:
$\begin{matrix} α_{2} = f (RI, GSI) \times α_{1} & (11) \\ f (RI, GSI) = {\begin{matrix} 0, if RI < {RI}_{T} or GSI > {GSI}_{T} \\ RI, if RI \geq {RI}_{T} and GSI \leq {GSI}_{T} and for k \leq C \\ 1 - RI, if RI \geq {RI}_{T} and GSI \leq {GSI}_{T} and for k > C \end{matrix} & (12) \end{matrix}$
C stands for a predetermined number of process runs. For a W2W control, C can be 25 for semiconductor industries.
Since the R2R controller 130 also can be a MA controller, a d-EWMA controller or a PID controller, a generic form of governing equations is provided as follows:
u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z) (13)
u _k+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,ŷ _k) (14)
G_2,i =f(RI_k,GSI_k)×G_1,i (15)
where G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller, if RI_k<RI_Tor GSI_k>GSI_T;
f(RI _k ,GSI _k)=RI _k, if RI_k≧RI_Tand GSI_k≦GSI_Tand k≦C;
f(RI _k ,GSI _k)=1−RI _k, if RI_k≧RI_Tand GSI_k≦GSI_Tand k>C;
wherein y_zrepresents the actual measurement value of the sampling workpiece which has been processed in the z^thprocess run; u_z+1represents the control action of the (z+1)^thprocess run when y_zis adopted; G_1,i,represents the controller gain used in the R2R controller when y_zis adopted, wherein i represents the number of the controller gains used in the R2R controller; ŷ_krepresents the virtual metrology value of the workpiece which has been processed in the k^thprocess run; u_k+1represents the control action of the (k+1)^thprocess run when ŷ_kis adopted; G_2,i,represents the controller gain used in the R2R controller when ŷ_kis adopted; RI_krepresents the reliance index (RI) of the k^thprocess run; RI_Trepresents a RI threshold value based on a maximal tolerable error limit defined by the errors of the virtual metrology values obtained from the conjecture model; GSI_krepresents the global similarity index (GSI) of the k^thprocess run; GSI_Trepresents a GSI threshold value defined by two to three times of the maximal global similarity indexes of the sets of historical process data; and C stands for a predetermined number of process runs.
The MA controller and the EWMA controller which are single-gain controllers; and the d-EWMA controller, and the PID controller which are multiple-gain controllers are described below.

MA Controller

The (z−1)^thrun control action, u_z+1, of an n-terms MA controller is derived by
$\begin{matrix} u_{z + 1} = \frac{{Tgt}_{z + 1} - {\tilde{η}}_{z + 1}}{A} & (16) \end{matrix}$
where A is a gain parameter (e.g., removal rate for chemical mechanical to polishing (CMP)) estimated for the system; Tgt_z+1is the target value of (z+1)^thrun; and {tilde over (η)}_z+1is the model offset or disturbance of the (z+1)^thrun. {tilde over (η)}_z+1of the n-terms MA controller is expressed as:
$\begin{matrix} \begin{matrix} {\tilde{η}}_{z + 1} = \frac{1}{n} (y_{z} - {Au}_{z}) + \frac{1}{n} (y_{z - 1} - {Au}_{z - 1}) + \dots + \frac{1}{n} (y_{z - (n - 1)} - {Au}_{z - (n - 1)}) \\ = \frac{1}{n} \sum_{i = 1}^{n} (y_{z - (i - 1)} - {Au}_{z - (i - 1)}) \\ = \frac{1}{n} (1 + q^{- 1} + q^{- 2} + \dots + q^{- (n - 1)}) (y_{z} - {Au}_{z}) \\ = M_{1} h_{MA} (q) (y_{z} - {Au}_{z}) \end{matrix} & (17) \end{matrix}$
where y_zrepresents the actual measurement value of the z^thrun control output; q represents the delay operator, i.e. q⁻¹y_z=y_z−1; M₁=1/n is the controller gain; and
h _MA(q)=(1+q ⁻¹ + . . . +q ⁻⁽ⁿ⁻¹⁾) (18)
Then, from equation (16),
$\begin{matrix} \begin{matrix} u_{z + 1} = \frac{{TgT}_{z + 1} - {\tilde{η}}_{z + 1}}{A} \\ = \frac{{Tgt}_{z + 1} - M_{1} h_{MA} (q) (y_{z} - {Au}_{z})}{A} \\ = g_{MA} (M_{1}, y_{z}) \end{matrix} & (19) \end{matrix}$
In conclusion, the (z+1)^thrun control action, u_z+1, of an n-terms MA controller can be expressed as a function of the actual measurement value of the z^thcontrol output, y_z, and the controller gain, M₁.

EWMA Controller

The (z+1)^thrun control action, u_z+1, of an EWMA controller can also be expressed as equation (16).
For the EWMA controller, {tilde over (η)}_Z+1is derived below.
$\begin{matrix} \begin{matrix} {\tilde{η}}_{z + 1} = α_{1} (y_{z} - {Au}_{z}) + (1 - α_{1}) {\tilde{η}}_{z} \\ = α_{1} (y_{z} - {Au}_{z}) + α_{1} (1 - α_{1}) (y_{z - 1} - {Au}_{z - 1}) + \\ {α_{1} (1 - α_{1})}^{2} (y_{z - 2} - {Au}_{z - 2}) + \dots + \\ {α_{1} (1 - α_{1})}^{i} (y_{z - i} - {Au}_{z - i}) + \dots + \\ {α_{1} (1 - α_{1})}^{z - 1} (y_{1} - {Au}_{1}) + {(1 - α_{1})}^{z} {\tilde{η}}_{0} \\ = \sum_{i = 1}^{z} {α_{1} (1 - α_{1})}^{z - i} (y_{i} - {Au}_{i}), \\ for initial condition {\tilde{η}}_{0} = 0. \end{matrix} & (20) \\ Let c_{0} = α_{1} c_{1} = α_{1} (1 - α_{1}) c_{2} = {α_{1} (1 - α_{1})}^{2} ⋮ c_{i} = {α_{i} (1 - α_{1})}^{i} ⋮ c_{z - 1} = {α_{1} (1 - α_{1})}^{z - 1} then \\ \begin{matrix} {\tilde{η}}_{z + 1} = (c_{0} + c_{1} \cdot q^{- 1} + \dots + c_{i} \cdot q^{- i} + \dots + c_{z - 1} \cdot q^{- (z - 1)}) \\ (y_{z} - {Au}_{z}) \\ = α_{1} h_{EWMA} (α_{1}, q) (y_{z} - {Au}_{z}) \end{matrix} and & (21) \\ h_{EWMA} (α_{1}, q) = [1 + (1 - α_{1}) q^{- 1} + {(1 - α_{1})}^{2} q^{- 2} + \dots + {(1 - α_{1})}^{z - 1} q^{- (z - 1)}] & (22) \\ \begin{matrix} u_{z + 1} = \frac{{Tgt}_{z + 1} - {\tilde{η}}_{z + 1}}{A} \\ = \frac{{Tgt}_{z + 1} - α_{1} h_{EWMA} (α_{1}, q) (y_{z} - {Au}_{z})}{A} \\ = g_{EWMA} (α_{1}, y_{z}) \end{matrix} & (23) \end{matrix}$
In conclusion, the (z+1)^thrun control action, u_z+1, of an EWMA controller can be expressed as a function of the actual measurement value of the z^thrun control output, y_z, and the controller gain, α₁.
d-EWMA Controller
The (z+1)^thrun control action, u_z+1, a d-EWMA controller is expressed as:
$\begin{matrix} u_{z + 1} = \frac{{Tgt}_{z + 1} - {\tilde{η}}_{z + 1} - {\tilde{ρ}}_{z + 1}}{A} & (24) \end{matrix}$
Referring to equations (20), (21), and (22), {tilde over (η)}_Z+1can be expressed as:
$\begin{matrix} \begin{matrix} {\tilde{η}}_{z + 1} = α_{1, 1} (y_{z} - {Au}_{z}) + (1 - α_{1, 1}) {\tilde{η}}_{z} \\ = α_{1, 1} h_{EWMA} (α_{1, 1}, q) (y_{z} - {Au}_{z}) \end{matrix} & (25) \end{matrix}$
Similarly, {tilde over (ρ)}_Z+, is derived as:
$\begin{matrix} \begin{matrix} {\tilde{ρ}}_{z + 1} = α_{1, 2} (y_{z} - {Au}_{z} - {\tilde{η}}_{z}) + (1 - α_{1, 2}) {\tilde{ρ}}_{z} \\ = α_{1, 2} h_{EWMA} (α_{1, 2}, q) (y_{z} - {Au}_{z} - {\tilde{η}}_{z}) . \end{matrix} & (26) \end{matrix}$
Finally, u_z+1can be expressed as:
$\begin{matrix} u_{z + 1} = \frac{{Tgt}_{z + 1} - {\tilde{η}}_{z + 1} - {\tilde{ρ}}_{z + 1}}{A} = g_{d - EWMA} (α_{1, 1}, α_{1, 2}, y_{z}) & (27) \end{matrix}$
In conclusion, the (z+1)^thrun control action, u_z+1, of a d-EWMA controller can be expressed as a function of the actual measurement value of the z^thrun control output, y_z, and the controller gains, α_1.1and α_1.2.

PID Controller

The (z+1)^thrun control action, u_z+1, of a PID controller is expressed as:
$\begin{matrix} \begin{matrix} u_{z + 1} = K_{1, P} (y_{z} - {Tgt}_{z}) - K_{1, I} \frac{1}{1 - q^{- 1}} (y_{z} - {Tgt}_{z}) - \\ K_{1, D} (1 - q^{- 1}) (y_{z} - {Tgt}_{z}) \\ = g_{PID} (K_{1, P}, K_{1, I}, K_{1, D}, y_{z}) \end{matrix} & (28) \end{matrix}$
In conclusion, the (z+1)^thrun control action, u_z+1, of a PID controller can be expressed as a function of the actual measurement value of the z^thrun control output, y_z, and the controller gains, K_1,P, K_1,I, and K_1,D.
Observing equations (19), (23), (27), and (28), a generic form of the (z+1)^thrun control action, u_z+1, of the MA, EWMA, d-EWMA, and PID R2R controller can be generated as a function of the actual measurement value of the z^thrun control output, y_z, and the controller gains, G_1,1, G_1,2, . . . , and G_1,i, where i represented the number of gains existed in the controller.
u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z) (29)
For the MA case, i=1 and G_1,1=M₁; for EWMA, i=1 and G_1,1=α₁; for d-EWMA, i=2, G_1,1=α_1,1and G_1,2=α_1,2; for PID, i=3, G_1,1=K_1,P, G_1,2=K_1,I, and G_1,3=K_1,D. In fact, equation (29) has been mentioned in equation (13).
When VM is utilized, y₂will be replaced by ŷ_kand the controller gains will be changed to G_2,1, G_2,2, . . . , and G_2,i, where i represented the number of gains existed in the controller. Therefore, by utilizing VM, the generic form of the (k+1)^thrun control action, u_k+1, is
u _k+1 =g(G _2,1 ,G _2,2 , . . . ,G _2,i,ŷ_k) (30)
For the MA case, i=1 and G_2,1=M₂; for EWMA, i=1 and G_2,1=α₂; for d-EWMA, i=2, G_2,1=α_2,1and G_2,2=α_2,2for PID, i=3, G_2,1=K_2,P, G_2,2=K_2,I, and G_2,3=K_2,D. In fact, equation (30) has been mentioned in equation (14).
When VM is adopted as the feedback of the R2R controller, VM's accompanying RI/GSI can be used to tune the controller gains as shown below:
G _2,i =f(RI,GSI)×G _1,i (31)
In fact, equation (31) has been mentioned in equation (15).
Specifically, for the MA case:
M ₂ =f _MA(RI,GSI)×M ₁ (32)
For the EWMA case:
α₂ =f _EWMA(RI, GSI)×α₁ (33)
For the d-EWMA case:
α_2,1 =f _α1(RI,GSI)×α_1,1
α_2,2 =f _α2(RI,GSI)×α_1.2 (34)
For the PID case:
K _2,P =f _P(RI,GSI)×K _1.P
K _2,I =f _I(RI,GSI)×K _1,I
K _2,D =f _D(RI,GSI)×K _1,D (35)
In conclusion, all of the G_1,Icontroller gains may be assigned as constants or tuned by an adaptive scheme or function. When the actual measurement values (y_z) are adopted, G_1,iwill be designed and assigned accordingly. After G_1,iare assigned and if the VM values (ŷ_k) are adopted to replace y_z, then the corresponding G_2,1gains can be designed and assigned as shown in equations (31)-(35).
Equations (31)-(35) are valid only when RI and GSI are good enough; in other words, RI should be greater than RI_Tand GSI should be smaller than GSI_T. If RI<RI_Tor GSI>GSI_T, this VM value cannot be adopted for tuning the R2R controller gain. In conclusion, if RI<RI_Tor GSI>GSI_T, then
for the MA case, set {tilde over (η)}_k+1={tilde over (η)}_k, i.e. ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller;
for the EWMA case, set {tilde over (η)}_k+1={tilde over (η)}_kor α₂=0 (i.e. G_2,i=0);
for the d-EWMA case, set {tilde over (η)}_k+1={tilde over (η)}_kand ρ _k+1={tilde over (ρ)}_k; or set α_2,1=α_2,2=0 (i.e. G_2,i=0);
for the PID case, set u_k+1=u_k, i.e. ŷ_k−1but not ŷ_kis adopted for tuning the R2R controller.
The following presents the algorithms related to the RI and explains their operating procedures.

Reliance Index (RI)

Referring to Table 1, n sets of historical data are assumed to be collected, including process data (X_i,i=1, 2, . . . ,n) and the corresponding actual measurement values (y_i,i=1, 2, . . . , n), where each set of process data contains p individual parameters (from parameter 1 to parameter p), namely X_i=[x_i,1, x_i,2, . . . , x_i,p]^T. Additionally, (m-n) sets of process data in actual production were also collected, but no actual measurement values are available besides y_n+1. That is, only the first among (m-n) pieces of the products is selected and actually measured. In the current manufacturing practice, the actual measurement value y_n+1obtained is used to infer and evaluate the quality of the (m-n−1) pieces of the products.

TABLE 1

					Actual
Sample	Parameter	Parameter		Parameter	Measurement
Data Set
	1	2		p	Value

1	x_{1, 1}	x_{1, 2}	. . .	x_{1, p}	y₁
2	x_{2, 1}	x_{2, 2}	. . .	x_{2, p}	y₂
. . .	. . .	. . .	. . .	. . .	. . .
n	x_{n, 1}	x_{n, 2}	. . .	x_{n, p}	y_n
n + 1	x_{n+1, 1}	x_{n+1, 2}	. . .	x_{n+1, p}	y_n+1
n + 2	x_{n+2, 1}	x_{n+2, 2}	. . .	x_{n+2, p}	Zip
. . .	. . .	. . .	. . .	. . .	. . .
m	x_{m, 1}	x_{m, 2}	. . .	x_{m, p}	Zip

As shown in Table 1, y₁, y₂, . . . , y_nare historical measurement values, and y_n+1is the actual measurement value of the first piece of the products being manufactured. Generally, a set of actual measurement values (y_i,i=1, 2, . . . ,n) is a normal distribution with mean μ and standard deviation σ, namely y_i˜N(μ, σ²).
All the actual measurement values can be standardized in terms of the mean and standard deviation of the sample set (y_i,i=1, 2, . . . , n). Their standardized values (also called z scores) Z_y ₁, Z_y ₂, . . . , Z_y _nare thus derived, where each z score has mean zero and standard deviation one, namely Z_y _i˜N(0, 1). Regarding the actual measurement data, a corresponding Z_y _iclose to 0 indicates that the actual measurement value approaches the central value of the specification. The equations for standardization are listed as follows:
$\begin{matrix} Z_{y_{i}} = \frac{y_{i} - \overline{y}}{σ_{y}}, i = 1, 2, \dots, n & (36) \\ \overline{y} = \frac{1}{n} (y_{1} + y_{2} + \dots + y_{n}) & (37) \\ σ_{y} = \sqrt{\frac{1}{n - 1} [{(y_{1} - \overline{y})}^{2} + {(y_{2} - \overline{y})}^{2} + \dots + {(y_{n} - \overline{y})}^{2}]} & (38) \end{matrix}$
wherein y_iis the i-th actual measurement value,

- Z_y _iis the standardized i-th actual measurement value,
- y is the mean of all the actual measurement values, and
- σ_yis the standard deviation of all the actual measurement values.

The explanation herein adopts a neural—networks (NN) algorithm as the conjecture algorithm for establishing the conjecture model performing virtual measurement, and uses such as a multi-regression (MR) algorithm to be the reference algorithm for establishing the reference model that serves as a comparison base for the conjecture model. However, the present invention can also apply other algorithms to be the conjecture algorithm or the reference algorithm, provided the reference algorithm differs from the conjecture algorithm, such as a support-vector-regression (SVR) algorithm, a partial-least-squares regression (PLSR) algorithm, a Gaussian-process-regression (GPR) algorithm or other related algorithms, and thus the present invention is not limited thereto.
When the NN and MR algorithms are utilized, if their convergence conditions both are that SSE (Sum of Square Error) is minimized with n→∞, their standardized predictive measurement values (defined as
$Z_{y_{N_{i}}} and Z_{y_{r_{i}}}$
respectively) should be the same as the standardized actual measurement value Z_y _i. Restated, when n→∞,
$Z_{y_{i}} = Z_{y_{N_{i}}} = Z_{y_{ri}}$
all represent the standardized actual measurement value, but they have different names due to having different purposes and different estimating models. Hence,
$Z_{y_{N_{i}}} ~ N (μ_{Z_{y_{i}}}, σ_{Z_{y}}^{2}) and Z_{y_{r_{i}}} ~ N (μ_{Z_{y_{i}}}, σ_{Z_{y}}^{2})$
indicate that Z_y _Niand Z_y _ri, share the same statistical distribution. However, owing to the existence of different estimating models, the estimations of mean and standard deviation differ between those two prediction algorithms. Namely the standardized mean-estimating equation ({circumflex over (μ)}_Z _yi=Z_ŷ _Ni) and standard-deviation-estimating equation
$({\hat{σ}}_{z_{y}} = {\hat{σ}}_{z_{{\hat{y}}_{N}}})$
with respect to the NN conjecture model differ from the standardized mean-estimating equation ({circumflex over (μ)}_Z _yi=Z_ŷ _ri) and standard-deviation-estimating equation
$({\hat{σ}}_{z_{y}} = {\hat{σ}}_{z_{{\hat{y}}_{r}}})$
with respect to the MR reference model.
The RI is designed to gauge the reliance level of the virtual metrology value. The RI thus should consider the degree of similarity between the statistical distribution Z_ŷ _Niof the virtual metrology value and the statistical distribution Z_y _iof the actual measurement value. However, when virtual metrology is applied, no actual measurement value can be used to verify the trustworthiness of the virtual metrology value. (Notably, virtual metrology becomes unnecessary if actual measurement values are obtained.) Instead, the present invention adopts the statistical distribution Z_ŷ _riestimated by the reference algorithm which is such as the MR algorithm to replace Z_y _i. The reference algorithm also can be such as a support-vector-regression (SVR) is algorithm, a partial-least-squares regression (PLSR) algorithm, a Gaussian-process-regression (GPR) algorithm, or other related algorithms, and thus the present invention is not limited thereto.
Referring to FIG. 4A, FIG. 4A is a schematic diagram for explaining the reliance index (RI) according to the preferred embodiment of the present invention. The RI of the present invention is defined as the intersection-area value (overlap area A) between the statistical distribution Z_ŷ _Niof the virtual metrology value from the conjecture model (built by such as the NN algorithm) and the statistical distribution
$Z_{{\hat{y}}_{r_{i}}}$
of the reference prediction value from the reference model (built by such as the MR algorithm). As such, the RI equation is listed below:
$\begin{matrix} RI = 2 \int_{\frac{Z_{{\hat{y}}_{N_{i}}} + Z_{{\hat{y}}_{ri}}}{2}}^{\infty} \frac{1}{\sqrt{2 π} σ} e^{- \frac{1}{2} {(\frac{x - μ}{σ})}^{2}} \partial x with μ = Z_{{\hat{y}}_{N_{i}}} if Z_{{\hat{y}}_{Ni}} < Z_{{\hat{y}}_{ri}} μ = Z_{{\hat{y}}_{r_{i}}} if Z_{{\hat{y}}_{ri}} < Z_{{\hat{y}}_{Ni}} & (39) \end{matrix}$
and σ is set to be 1.
The RI increases with increasing overlap area A. This phenomenon indicates that the result obtained using the conjecture model is closer to that obtained from the reference model, and thus the corresponding virtual metrology value is more reliable. Otherwise, the reliability of the corresponding measurement value reduces with decreasing RI. When the distribution Z_ŷ _Niestimated from Z_y _Niis fully overlapped with the distribution Z_ŷ _riestimated from Z_y _ri, then according to the distribution theory of statistics, the RI value equals 1; and, when those two distributions are almost separate, the RI value approaches 0.
Hereinafter, the method for calculating the statistical distribution of the virtual metrology values (Z_ŷ _Niand {circumflex over (σ)}_Z _ŷN) from the conjecture model is explained.
In the NN conjecture model, if the convergence condition is to minimize SSE, then it can be assumed that “for given Z_x _i,j, Z_y _Niis the distribution with mean μ_Z _yiand standard deviation σ_Z _y”, namely for given Z_x _i,j, Z_y _Ni˜N(μ_Z _yi,σ_Z _y ²), where the NN estimating equation of σ_Z _y ²is {circumflex over (σ)}_Z _y ²={circumflex over (σ)}_Z _ŷN. and the NN estimating equation of σ_Z _y ²is {circumflex over (σ)}_Z _y ²={circumflex over (σ)}_Z _ŷN ².
Before the NN conjecture model is constructed, the process data must be standardized. The equations for standardizing the process data are presented below:
$\begin{matrix} Z_{x_{i, j}} = \frac{x_{i, j} - {\overline{x}}_{j}}{σ_{x_{j}}}, i = 1, 2, \dots, n, n + 1, \dots, m; j = 1, 2, \dots, p & (40) \\ {\overline{x}}_{j} = \frac{1}{n} (x_{1, j} + x_{2, j} + \dots + x_{n, j}) & (41) \\ σ_{x_{j}} = \sqrt{\frac{1}{n - 1} [{(x_{1, j} - {\overline{x}}_{j})}^{2} + {(x_{2, j} - {\overline{x}}_{j})}^{2} + \dots + {(x_{n, j} - {\overline{x}}_{j})}^{2}]} & (42) \end{matrix}$
wherein x_i,jis the j-th process parameter in the i-th set of process data,

- Z_x _i,jis the standardized j-th process parameter in the i-th set of process data;
- x _jis the mean of the j-th process data;
- σ_x _jis the standard deviation of the j-th process data.

The n sets of standardized process data (Z_x _i,j,i=1,2, . . . ,n; j=1,2, . . . ,p) and the n standardized actual measurement values (Z_y _i,i=1,2, . . . ,n) are utilized to build the NN conjecture model. The m sets of standardized process data (Z_x _i,j,i=1,2, . . . ,m; j=1,2, . . . ,p) are then inputted into the NN conjecture model to obtain the corresponding standardized virtual metrology values:
$Z_{{\hat{y}}_{N_{1}}}, Z_{{\hat{y}}_{N_{2}}}, \dots, Z_{{\hat{y}}_{N_{n}}}, Z_{{\hat{y}}_{N_{n + 1}}}, \dots, Z_{{\hat{y}}_{N_{m}}} .$
Accordingly, the estimated value of μ_Z _yi(i.e. {circumflex over (μ)}_Z _yi=Z_ŷ _Ni) and the estimated value of σ_Z _y(i.e. σ_Z _y={circumflex over (σ)}_Z _ŷN) can be computed as follows:
$\begin{matrix} {\hat{μ}}_{Z_{y_{i}}} = Z_{{\hat{y}}_{N_{i}}}, i = 1, 2, \dots, n, n + 1, \dots, m & (43) \\ {\hat{σ}}_{Z_{\hat{y} N}} = \sqrt{\frac{1}{n - 1} [\begin{matrix} {(Z_{{\hat{y}}_{N_{1}}} - {\overline{Z}}_{{\hat{y}}_{N}})}^{2} + {(Z_{{\hat{y}}_{N_{2}}} - {\overline{Z}}_{{\hat{y}}_{N}})}^{2} + \dots + \\ {(Z_{{\hat{y}}_{N_{n}}} - {\overline{Z}}_{{\hat{y}}_{N}})}^{2} \end{matrix}]} & (44) \\ {\overline{Z}}_{{\hat{y}}_{N}} = \frac{1}{n} (Z_{{\hat{y}}_{N_{1}}} + Z_{{\hat{y}}_{N_{2}}} + \dots + Z_{{\hat{y}}_{N_{n}}}) & (45) \end{matrix}$
wherein Z_ŷ _Nis the mean of the standardized virtual metrology values.
Hereinafter, the method for calculating the reference predication values (Z_ŷ _riand {circumflex over (σ)}_Z _ŷr) from the MR model is explained.
The basic assumption of the MR is that “for given Z_y _i,j, Z_y _riis the distribution with mean μ_Z _yiand standard deviation σ_Z _y”, namely for given Z_x _i,j, Z_y _ri˜N(μ_Z _yi, σ_Z _y ²), wherein the MR estimating equation of
$μ_{Z_{y_{i}}}$
is {circumflex over (μ)}_Z _yi=Z_ŷ _ri, and the MR estimating equation of σ_Z _y ²is {circumflex over (σ)}_Z _y ²={circumflex over (σ)}_Z _ŷr ².
To obtain the MR relationship between the n sets of standardized process data (Z_x _i,ji=1,2, . . . ,n; j=1,2, . . . , p) and the n standardized actual measurement values, (Z_y _i=1,2, . . . ,n), the weighting factors β_r=[β_r0,β_r1,β_r2, . . . ,β_rp]^Tcorresponding to those p parameters must be defined by using the MR analysis. The relationship between Z_y _iand Z_x _i,jthus is constructed as follows:
$\begin{matrix} β_{r 0} + β_{r 1} Z_{x_{1, 1}} + β_{r 2} Z_{x_{1, 2}} + \dots + β_{rp} Z_{x_{1, p}} = Z_{y_{1}} β_{r 0} + β_{r 1} Z_{x_{2, 1}} + β_{r 2} Z_{x_{2, 2}} + \dots + β_{rp} Z_{x_{2, p}} = Z_{y_{2}} \dots β_{r 0} + β_{r 1} Z_{x_{n, 1}} + β_{r 2} Z_{x_{n, 2}} + \dots + β_{rp} Z_{x_{n, p}} = Z_{y_{n}} Let & (46) \\ Z_{y} = (\begin{matrix} Z_{y_{1}} \\ Z_{y_{2}} \\ ⋮ \\ Z_{y_{n}} \end{matrix}) and & (47) \\ Z_{x} = (\begin{matrix} 1 & Z_{x_{1, 1}} & \dots & Z_{x_{1, p}} \\ 1 & Z_{x_{2, 1}} & \dots & Z_{x_{2, p}} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ 1 & Z_{x_{n, 1}} & \dots & Z_{x_{n, p}} \end{matrix}) & (48) \end{matrix}$
The least square method can obtain the estimating equation of β_r, {circumflex over (β)}_r=[{circumflex over (β)}_r0,{circumflex over (β)}_r1,{circumflex over (β)}_r2, . . . ,{circumflex over (β)}_rp]^Tas
{circumflex over (β)}_r=(Z _x ^T Z _x)⁻¹ Z _x ^T Z _y (49)
Therefore, the MR reference model can be obtained as
Z_ŷr _i={circumflex over (β)}_r0+{circumflex over (β)}_r1 Z _x _i,1+{circumflex over (β)}_r2 Z _x _i,2+ . . . +{circumflex over (β)}_rp Z _x _i,p
i=1,2, . . . ,n,n+1, . . . ,m (50)
Hence, during the conjecture phase, after inputting a set of process data, its MR estimating value Z_ŷ _ricorresponding thereto can be obtained via equation (50). The MR estimating equation of the standard deviation σ_Z _yis {circumflex over (σ)}_Z _ŷrwith
$\begin{matrix} {\hat{σ}}_{Z_{\hat{y} r}} = \sqrt{\frac{1}{n - 1} [\begin{matrix} {(Z_{{\hat{y}}_{r_{1}}} - {\overline{Z}}_{{\hat{y}}_{r}})}^{2} + {(Z_{{\hat{y}}_{r_{2}}} - {\overline{Z}}_{{\hat{y}}_{r}})}^{2} + \dots + \\ {(Z_{{\hat{y}}_{r_{n}}} - {\overline{Z}}_{{\hat{y}}_{r}})}^{2} \end{matrix}]} & (51) \\ {\overline{Z}}_{{\hat{y}}_{r}} = \frac{1}{n} (Z_{{\hat{y}}_{r_{1}}} + Z_{{\hat{y}}_{r_{2}}} + \dots + Z_{{\hat{y}}_{r_{n}}}) & (52) \end{matrix}$
After obtaining the NN estimating equations (Z_ŷ _Niand {circumflex over (σ)}_Z _ŷN) and the MR estimating equations (Z_ŷ _riand their normal distribution curves can be depicted, as illustrated in FIG. 4A. Thus, the RI of each virtual metrology value can be derived by calculating the intersection area value (overlap area A).
After obtaining the RI, the RI threshold value (RI_T) must be defined. If RI>RI_T, then the reliance level of the virtual metrology value is acceptable. A systematic approach for determining the RI_Tis described below.
Before determining the RI_T, it is necessary to define a maximal tolerable error limit (E_L). The error of the virtual metrology value is an absolute percentage of the difference between the actual measurement value y_iand ŷ_Niobtained from the NN conjecture model divided by the mean of all the actual measurement values, y, namely
$\begin{matrix} {Error}_{i} = \langle \frac{y_{i} - {\hat{y}}_{Ni}}{\overline{y}} \rangle \times 100 % & (53) \end{matrix}$
The E_Lcan then be specified based on the error defined in equation (53) and the accuracy specification of virtual metrology (VM). Consequently, RI_Tis defined as the RI value corresponding to the E_L, as shown in FIG. 4B. That is
$\begin{matrix} {RI}_{T} = 2 \int_{Z_{Center}}^{\infty} \frac{1}{\sqrt{2 π} σ} e^{- \frac{t}{2} {(\frac{x - μ}{σ})}^{2}} \partial x & (54) \end{matrix}$
with μ and σ defined in equation (39) and
Z _Center =Z _ŷ _Ni +[ y×(E _L/2)]/σ_y (55)
where σ_yis specified in equation (38).
The following presents the algorithms related to the GSI and explains their operating procedures.

Global Similarity Indexes (GSI)

When virtual metrology is applied, no actual measurement value is available to verify the accuracy of the virtual metrology value. Therefore, instead of the standardized actual measurement value Z_y _ithe standardized MR prediction value Z_ŷ _riis adopted to calculate the RI. This substitution may cause inevitable gauging errors in the RI. To compensate for this unavoidable substitution, a global similarity index (GSI) is provided to help the RI gauge the reliance level of virtual metrology and identifying the key process parameters with large deviations (z score values).
The GSI assesses the degree of similarity between any set of process data and the model set of process data. This model set is derived from all of the sets of historical process data used for building the conjecture model.
The present invention may utilize a statistical distance measure, such as Mahalanobis distance, to quantify the degree of similarity. Mahalanobis distance is a distance measure introduced by P.C. Mahalanobis in 1936. This measure is based on correlation between variables to identify and analyze different patterns of sample sets. Mahalanobis distance is a useful way of determining similarity of an unknown sample set to a known one. This method considers the correlation of the data set and is scale-invariant, namely it is not dependent on the scale of measurements. If the data set has high similarity, the calculated Mahalanobis distance calculated will be relatively small.
The present invention uses the calculated GSI (applying Mahalanobis distance) size to determine whether the newly input set of process data is similar to the model set of process data. If the calculated GSI is small, the newly input set is relatively similar to the model set. Thus the virtual metrology value of the newly input (high-similarity) set is relatively accurate. On the contrary, if the calculated GSI is too large, the newly input set is somewhat different from the model set. Consequently, the virtual metrology value estimated in accordance with the newly input (low-similarity) set has low reliance level in terms of accuracy.
The equations to calculate the standardized process data Z_x _i,jof the conjecture model are shown in equations (40), (41) and (42). At first, the model set of the process parameters is defined as X_M=[x_M,1,x_M,2, . . . ,x_M,p]^T, where x_M,jequals x _j,j=1,2, . . . ,p, so that each element in the model set after standardization (also denoted as the standardized model parameter, Z_M,j) has a value of 0. Restated, all of the elements in Z_M=[Z_M,1,Z_M,2, . . . ,Z_M,p]^Tare 0. Thereafter, the correlation coefficients between the standardized model parameters are calculated.
Assuming that the correlation coefficient between the s-th parameter and the t-th parameter is r_stand that there are k sets of data, then
$\begin{matrix} r_{st} = \frac{1}{k - 1} \sum_{l = 1}^{k} z_{s 1} \cdot z_{t 1} = \frac{1}{k - 1} (z_{s 1} \cdot z_{t 1} + z_{s 2} \cdot z_{t 2} + \dots + z_{sk} \cdot z_{tk}) & (56) \end{matrix}$
After calculating the correlation coefficients between the standardized model parameters, the matrix of correlation coefficients can be obtained as
$\begin{matrix} R = [\begin{matrix} 1 & r_{12} & \dots & r_{1 p} \\ r_{21} & 1 & \dots & r_{2 p} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ r_{p 1} & r_{p 2} & \dots & 1 \end{matrix}] & (57) \end{matrix}$
Assuming that the inverse matrix (R⁻¹) of R is defined as A, then
$\begin{matrix} A = R^{- 1} = [\begin{matrix} a_{11} & a_{12} & \dots & a_{1 p} \\ a_{21} & a_{22} & \dots & a_{2 p} \\ \dots & \dots & \dots & \dots \\ a_{p 1} & a_{p 2} & \dots & a_{pp} \end{matrix}] & (58) \end{matrix}$
Hence, the equation for calculating the Mahalanobis distance (D_λ ²) between the standardized λ-th set process data (Z_λ) and the standardized model set process data (Z_M) is as follows.
$\begin{matrix} \begin{matrix} D_{λ}^{} = {(Z_{λ} - Z_{M})}^{T} R^{- 1} (Z_{λ} - Z_{M}) \\ = Z_{λ}^{T} R^{- 1} Z_{λ} \end{matrix} & (59) \end{matrix}$
Finally, we have
$\begin{matrix} D_{λ}^{} = \sum_{j = 1}^{p} \sum_{i = 1}^{p} a_{ij} z_{i λ} z_{j λ} & (60) \end{matrix}$
The GSI of the standardized λ-th set process data is, then, equal to D_λ ²/p.
After obtaining the GSI, the GSI threshold (GSI_T) should be defined. Generally, the default GSI_Tis assigned to be two to three times the maximal GSI, (the subscript “a” stands for each historical set during the training phase).
Referring to FIG. 5, FIG. 5 is a schematic flow chart showing an APC method according to an embodiment of the present invention. In the APC method, step 200 is performed for obtaining a plurality of sets of historical process data used by a process tool for processing a plurality of historical workpieces. Step 210 is performed for obtaining a plurality of historical measurement data of the historical workpieces measured by a metrology tool, wherein the historical measurement values are the measurement values of the historical workpieces which are manufactured in accordance with the sets of historical process data stated in Step 200, respectively. Step 220 is performed for establishing a conjecture model in accordance with a conjecture algorithm by using the sets of historical process data and the historical measurement values; establishing a reference model in accordance with a reference algorithm by using the sets of historical process data and their corresponding historical measurement values and establishing a statistical distance model in accordance with a statistical distance algorithm by using the sets of historical process data. Step 230 is performed for enabling a run-to-run (R2R) controller to control the process tool to perform the process runs in accordance with the aforementioned equations (13)-(15).
The aforementioned embodiments can be provided as a computer program product, which may include a machine-readable medium on which instructions are stored for programming a computer (or other electronic devices) to perform a process based on the embodiments of the present invention. The machine-readable medium can be, but is not limited to, a floppy diskette, an optical disk, a compact disk-read-only memory (CD-ROM), a magneto-optical disk, a read-only memory (ROM), a random access memory (RAM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic or optical card, a flash memory, or another type of media/machine-readable medium suitable for storing electronic instructions. Moreover, the embodiments of the present invention also can be downloaded as a computer program product, which may be transferred from a remote computer to a requesting computer by using data signals via a communication link (such as a network connection or the like).
Hereinafter, illustrative examples are provided and compared for explaining that the embodiment of the present invention is useful and advantageous.
The W2W control of a CMP tool with a periodic maintenance (PM) cycle being 600 pieces (pcs) of wafers is selected as the illustrative example for evaluation and comparisons. The simulation conditions and scenarios are listed as follows:
1. y_kis the actual removal amount measured from the metrology tool and PostY_kis the actual post CMP thickness of run k. The specification of PostY_kis 2800±150 Angstrom (Å) with 2800 being the target value denoted by Tgt_PostY. Therefore, we have
PostY_k=PreY_k −y _k (61)
with
PostY_k=ARR_k *u _k (62)
where ARR_kis the actual removal rate of run k and u_krepresents the polish time in this example.
The well-known Preston equation, empirically found from the experiment of the glass polishing in 1927, has been proposed to predict the material removal rate of CMP. According to the Preston equation, the material removal rate is affected by the contact pressure (also denoted as tool stress) distribution at contact point, magnitude of the relative velocity (also denoted as tool rotation speed) at contact point between wafer and polishing pad, and constant representing the effect of the other remaining parameters including the slurry fluid speed, pad property, and so on. Therefore, ARR_kis simulated by:
$\begin{matrix} {ARR}_{k} = (\begin{matrix} A_{k} \times (\frac{Stress 1 + Stress 2}{1000}) \times (\frac{Rotspd 1 + Rotspd 2}{100}) \times \\ (\frac{Sfuspd 1 + Sfuspd 2}{100}) \end{matrix}) + (PM 1 + PM 2) + Error & (63) \end{matrix}$
The meanings of Stress1, Stress2, Rotspd1, Rotspd2, Sfuspd1, Sfuspd2, PM1, PM2, and Error are tabulated in Table 2. The A_kin equation (63) is the nominal removal rate, which is empirically simulated by a polynomial curve fitting of parts usage count between PMs (denoted by PU varying from 1 to 600):
A _k=(4×10⁻⁶)×(PU−1)³−(3.4×10⁻³)×(PU−1)²+(6.9×10⁻³)×(PU−1)+(1.202×10³) (64)
2. PostŶ_krepresents the predictive value of PostY_k, and then, from equations (61) and (62) we have
ŷ _k=A{circumflex over (R)}R_k *u _k (65)
PostŶ_k=PreY_k −ŷ _k=PreY_k−A{circumflex over (R)}R_k *u _k (66)
where
A{circumflex over (R)}R_k =f(Stress, Roupd, Sfuspd, PU,PU²,PU³) (67)
A{circumflex over (R)}R_kis the VM value of ARR_kwith Stress (=Stress1+Stress2), Rotspd (=Rotspd1+Rotspd2), Sfuspd (=Sfuspd1+Sfuspd2), PU, PU², PU³as the process parameters. The reason of adopting Stress, Rotspd, Sfuspd, PU, PU², and PU³as the process parameters is based on the Preston equation, equations (63) and (64). The setting values of the simulated process parameters are tabulated in Table 2.

TABLE 2

Simulation-Parameter Definitions and Setting Values

Setting Values

Abbreviation	Definition	Mean	VAR

Error	Random error represented by white	0	300
	noise
PM1	Error due to tool-parts' variation	0	100
	that was caused by periodic
	maintenance (PM)
PM2	Random disturbance of tool-parts'	0	6
	variation
Stress1	Tool stress error due to	1000	2000
	re-assembly during PM
Stress2	Random disturbance of tool stress	0	20
Rotspd1	Tool rotation-speed error due to	100	25
	re-assembly during PM
Rotspd2	Random disturbance of tool	0	1.2
	rotation speed
Sfuspd1	Slurry fluid-speed error due to	100	25
	re-assembly during PM
Sfuspd2	Random disturbance of slurry fluid	0	1.2
	speed
PreY_k	Pre-process (etching depth) value	3800	2500
	that affects the process result
	of run k

3. The k+1 run control action is derived by
$\begin{matrix} {Tgt}_{k + 1} = {PreY}_{k + 1} - {Tgt}_{PostY} & (68) \\ u_{k + 1} = \frac{{Tgt}_{k + 1} - {\tilde{η}}_{k + 1}}{A_{k + 1}} & (69) \end{matrix}$
4. When PostY_kis measured by an actual metrology tool, then
{tilde over (η)}_k+1=α₁(y _z −A _k u _k)+(1−α₁)η_k (70)
When PostY_kis conjectured or predicted by a VM system, then
$\begin{matrix} {\tilde{η}}_{k + 1} = α_{2, k} ({\hat{y}}_{k} - A_{k} u_{k}) + (1 - α_{2, k}) {\tilde{η}}_{k} with & (71) \\ α_{2, k} = f ({RI}_{k}, {GSI}_{k}) \times α_{1} where & (72) \\ f ({RI}_{k} {GSI}_{k}) = {\begin{matrix} 0, if {RI}_{k} < {RI}_{T} or {GSI}_{k} > {GSI}_{T} \\ \begin{matrix} {RI}_{k} if {RI}_{k} \geq {RI}_{T} and \\ {GSI}_{k} \leq {GSI}_{T} and for k \leq C \end{matrix} \\ \begin{matrix} 1 - {RI}_{k} if {RI}_{k} \geq {RI}_{T} and \\ {GSI}_{k} \leq {GSI}_{T} and for k > C \end{matrix} \end{matrix} & (73) \end{matrix}$
For this example, C=25.
5. 1 Lot=25 workpieces in which the 2^ndworkpiece being the sampling wafer.
6.
$\begin{matrix} Cpk (Process Capability) = \min {\begin{matrix} \frac{UCL - mean (PostY)}{3 \times std (PostY)}, \\ \frac{mean (PostY) - LCL}{3 \times std (PostY)} \end{matrix}} with UCL = 2950 and LCL = 2650. & (74) \\ 7. {MAPE}_{Process} = \frac{\sum_{i = 1}^{k} \langle ({PostY}_{i} - {Tgt}_{PostY}) {Tgt}_{PostY} \rangle}{k} \times 100 % & (75) \end{matrix}$
8. Extra random disturbances caused by Sfuspd2 with mean=0 and variance=0.36 are also added at Samples 50, 111, 179, 251, 349, and 503. In other words, the combined variances of Sfuspd2 at Samples 50, 111, 179, 251, 349, and 503 are 1.2+0.36=1.56. With these extra random disturbances, the to RI and/or GSI values may exceed their thresholds.
Five rounds with different random seeds are performed to evaluate and compare the performance. For each round, the simulation results of PreY_k, Tgt_k, A_k, and ARR_kfor k=1˜600 should be generated firstly based on the setting values shown in Table 2, equations (68), (64) and (63), respectively. Then, let α₁=0.35 and {tilde over (η)}₁=0 to calculate u₁as well as apply equations (62), (70), (69) and (61) to calculate y_k, η_k+1, u_k+1and PostY_k, respectively for k=1 and 2 for all of the five cases. As for k=3˜600, control schemes for those five cases are different and are described below:
Case 1: R2R with in-situ metrology
Let α₁=0.35. Apply equations (62), (70), (69) and (61) to calculate y_k, {tilde over (η)}_k+1, u_k+1, and PostY_k, respectively for k=3˜600.
Case 2: R2R+VM without RI
Let α₂=α₁=0.35. Apply equations (65), (71), (69), (66) and (61) to calculate ŷ_k, {tilde over (η)}_k+1, u_k+1, PostŶ_k, and PostY_k, respectively for k=3˜600.
Case 3: R2R+VM with RI
Let α₁=0.35. If RI<RI_Tor GSI>GSI_T, then let α₂=0; otherwise, let α_2,k=RI_k×α₁; as well as apply equations (65), (71), (69), (66) and (61) to calculate ŷ_k, {tilde over (η)}_k+1, u_k+1, PostŶ_k, and PostY_k, respectively for k=3˜600.
Case 4: R2R+VM with (1−RI)
Let α₁=0.35. If RI<RI_Tor GSI>GSI_T, then let α₂=0; otherwise, let α_2,k=(1−RI_k)×α₁; as well as apply equations (65), (71), (69), (66) and (61) to calculate ŷ_k, {tilde over (η)}_k+1, u_k+1, PostŶ_k, and PostY_k, respectively for k=3˜600.
Case 5: R2R+VM with RII(1−RI)
Let α₁=0.35. Apply the RII(1−RI) switching scheme as shown in equations (72) and (73) to set α₂; as well as apply equations (65), (71), (69), (66), and (61) to calculate ŷ_k, {tilde over (η)}_k+1, u_k+1, PostŶ_k, and PostY_k, respectively for k=3˜600.
Both Cpk (Process Capability Index) and MAPE_Process(Mean Absolute Percentage Error; as expressed in equations (74) and (75), respectively) are applied to evaluate and compare the performance of those 5 cases. The Cpk and MAPE_Processvalues of those 5 cases are tabulated in Tables 3 and 4, respectively.
Observing Tables 3 and 4 and treating Case 1 as the baseline, it is obvious that the performance of Case 2, which does not consider RI/GSI, is the worst. Case 3, which filters out the bad-quality PostŶ_k(VM) values and lets α₂=RI×α₁, is the most natural approach and has acceptable performance. The performance of Case 4, which filters out the bad-quality PostŶ_k(VM) values and lets α₂=(1−RI)×α₁, is better than that of Case 3 on average except for Round 1. Case 5, which filters out the bad-quality PostŶ_k(VM) values and applies the RII(1−RI) switching scheme shown in equation (73), fixes the problem of Case 4 in Round 1; and Case 5's performance is compatible with that of Case 1 (in-situ metrology).

TABLE 3

Cpk Values of 5-Cases APC Methods

Case 1: Insitu

Case 2: VM

Case 3: VM + RI

Case 4: VM + (1 − RI)

Case 5: VM + RI/(1 − RI)

Round	1~25	1~200	1~600	1~25	1~200	1~600	1~25	1~200	1~600	1~25	1~200	1~600	1~25	1~200	1~600

1	1.09	1.58	1.62	1.14	1.42	1.31	1.14	1.54	1.49	1.12	1.29	1.38	1.14	1.57	1.55
2	1.73	1.89	1.86	1.51	1.64	1.72	1.51	1.73	1.77	1.89	2.00	2.04	1.51	1.71	1.74
3	1.60	1.74	1.77	1.72	1.64	1.77	1.72	1.72	1.80	1.76	1.79	1.87	1.72	1.85	1.90
4	1.43	1.95	1.87	1.45	1.74	1.72	1.45	1.89	1.76	1.51	1.95	1.87	1.45	1.94	1.87
5	1.32	1.85	1.81	1.41	1.78	1.71	1.41	1.83	1.79	1.33	1.77	1.81	1.41	1.89	1.86
mean	1.43	1.80	1.80	1.45	1.64	1.65	1.45	1.74	1.72	1.52	1.76	1.80	1.45	1.79	1.78

TABLE 4

MAPEProcess Values of 5-Cases APC Methods

Case 1: Insitu

Case 2: VM

Case 3: VM + RI

Round	1~25	1~200	1~600	1~25	1~200	1~600	1~25	1~200	1~600

1	1.13%	0.86%	0.86%	1.35%	1.00%	1.07%	1.35%	0.94%	1.00%
2	0.85%	0.75%	0.76%	0.98%	0.87%	0.81%	0.97%	0.84%	0.79%
3	0.84%	0.85%	0.82%	0.94%	0.86%	0.81%	0.94%	0.83%	0.79%
4	0.93%	0.73%	0.76%	1.11%	0.83%	0.83%	1.11%	0.78%	0.82%
5	0.99%	0.75%	0.78%	1.05%	0.78%	0.82%	1.05%	0.77%	0.80%
mean	0.95%	0.79%	0.80%	1.08%	0.87%	0.87%	1.09%	0.83%	0.84%

Case 4: VM + (1 − RI)

Case 5: VM + RI/(1 − RI)

Round	1~25	1~200	1~600	1~25	1~200	1~600

1	2.52%	1.45%	1.15%	1.35%	0.98%	0.97%
2	0.80%	0.72%	0.69%	0.97%	1.05%	0.82%
3	1.03%	0.80%	0.76%	0.94%	0.77%	0.75%
4	1.07%	0.84%	0.79%	1.11%	0.84%	0.79%
5	1.41%	0.83%	0.78%	1.05%	0.77%	0.76%
mean	1.37%	0.93%	0.83%	1.09%	0.88%	0.82%

Simulation Results of Round 1 for those 5 cases are shown in FIG. 6A to is FIG. 6E, wherein the first 400 workpieces are shown in the figures. Due to the fact that extra random disturbances caused by Sfuspd2 with mean=0 and variance=0.36 are added at Samples 50, 111, 179, 251, 349, and 503, bad PostŶ_k(VM) values are generated and displayed in FIG. 6B. Those bad VM values as shown in FIG. 6B can be detected by RI and/or GSI.
The RI_Tand GSI_Tare set at 0.7 and 9, respectively in this example. The cases that RI<RI_Tand GSI>GSI_Tat Sample 50 of Round 1 as well as GSI>GSI_Tat Sample 349 of Round 1 are enlarged and depicted in FIG. 7 and FIG. 8, respectively.
As shown in FIG. 7, the Sample 50's PostŶ₅₀(VM) values of various cases are deviated with extra variance 0.36 adding to Sfuspd2. Owing to the fact that RI<RI_Tand GSI>GSI_T, those PostŶ₅₀values of Cases 3, 4, and 5 are filtered out by setting α₂=0, while the PostŶ₅₀value of Case 2 is still adopted to adjust the R2R controller gain with α₂=α₁=0.35. The effect of filtering out the to bad-quality PostŶ₅₀value is displayed in Sample 51, which shows that the PostY₅₁value of Case 2 is pulled down by the R2R controller since the PostŶ₅₀value at Sample 50 is too high. As for the other cases, Post Y₅₀and PostY₅₁have no much difference.
Observing FIG. 8, the Sample 349's PostŶ₃₄₉values of various cases are deviated with extra variance 0.36 adding to Sfuspd2, again. In this case, only GSI exceeds its threshold. By the same token, these bad PostY₃₄₉values are discarded in Cases 3, 4, and 5 yet not in Case 2. As such, improper R2R control result of Case 2, which generates a surge PostY₃₅₀, is shown in FIG. 8. The evidences displayed in FIG. 7 and FIG. 8 reveal that the result of adopting an unreliable VM value is worse than if no VM at all is utilized.
As mentioned above, α₂=RI×α₁when PostY_kis apart from the target value or production process is relatively unstable. On the contrary, if PostY_kis near the target or production process is relatively stable, then α₂=(1−RI)×α₁.
It will be apparent to those skilled in the art that various modifications and variations can be made to the structure of the present invention without departing from the scope or spirit of the invention. In view of the foregoing, it is intended that the present invention cover modifications and variations of this invention provided they fall within the scope of the following claims and their equivalents.

Claims

1. An advanced process control (APC) system, comprising:

a process tool for processing a plurality of historical workpieces in accordance with a plurality of sets of historical process data, and performing a plurality of process runs on a plurality of workpieces in accordance with a plurality of sets of process data;

a metrology tool for measuring said historical workpieces and a plurality of sampling workpieces selected from said workpieces, thereby providing a plurality of historical measurement data of said historical workpieces and a plurality of actual measurement values of said sampling workpieces which have been processed in said process runs;

a virtual metrology (VM) module for providing a plurality of virtual metrology values of said process runs by inputting said sets of process data into a conjecture model, wherein said conjecture model is built in accordance with a conjecture algorithm by using said sets of historical process data and said historical measurement values corresponding thereto, wherein said historical measurement values are the measurement values of said historical workpieces which are manufactured in accordance with said sets of historical process data, respectively;

a reliance index (RI) module for generating respective reliance indexes (RI) of said process runs, wherein each of said reliance indexes (RI) corresponding to the process run is generated by calculating the overlap area between the statistical distribution of the virtual metrology value of the workpiece and the statistical distribution of a reference prediction value of the workpiece, wherein said reference prediction value of the process run is generated by inputting the set of process data of the workpiece into a reference model, wherein said reference model is built in accordance with a reference algorithm by using said sets of historical process data and said historical measurement values corresponding thereto, and said conjecture algorithm is different from said reference algorithm, and the reliance index is higher when the overlap area is larger, representing that the reliance level of the virtual metrology value corresponding to the reliance index is higher;

a run-to-run (R2R) controller for controlling said process tool to perform said process runs in accordance with the following relationships:

u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z)

u _k+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,ŷ _k)

G _2,i =f(RI _k)×G _1,i;

where G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning said R2R controller, if R_k<RI_T;

f(RI _k)=RI _k, if RI_k≧RI_Tand k≦C;

f(RI _k)=1—RI _k, if RI_k≧RI_Tand k>C;

wherein y_zrepresents the actual measurement value of the sampling workpiece which has been processed in the z^thprocess run;

u_z+1represents the control action of the (z+1)^thprocess run when y_zis adopted;

G_1,i, represents the controller gain used in said R2R controller when y_zis adopted, wherein i represents the number of the controller gains used in said R2R controller;

ŷ_krepresents the virtual metrology value of the workpiece which has been processed in the k^thprocess run;

u_k+1represents the control action of the (k+1)^thprocess run when ŷ_kis adopted;

G_2,i, represents the controller gain used in said R2R controller when ŷ_kis adopted;

RI_krepresents the reliance index (RI) of the kth process run

RI_Trepresents the RI threshold value based on the maximal tolerable error limit defined by the errors of the virtual metrology values obtained from said conjecture model; and

C stands for a predetermined number of process runs.

2. The APC system as claimed in claim 1, wherein said conjecture algorithm and said reference algorithm are respectively selected from the group consisting of a multi-regression (MR) algorithm, a support-vector-regression (SVR) algorithm, a neural-networks (NN) algorithm, a partial-least-squares regression (PLSR) algorithm, and a Gaussian-process-regression (GPR) algorithm.

3. The APC system as claimed in claim 1, further comprising:

a global similarity index (GSI) module for generating respective global similarity indexes (GSI) of said process runs by inputting said sets of process data into a statistical distance model, wherein said statistical distance model is built in accordance with a statistical distance algorithm by using said sets of historical process data, wherein G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning said R2R controller, if GSI_k>GSI_T, where GSI_krepresents the global similarity index (GSI) of the kth process run; GSI_Trepresents the GSI threshold value defined by two to three times of the maximal global similarity indexes of said sets of historical process data.

4. The APC system as claimed in claim 3, wherein said statistical distance algorithm is a Mahalanobis distance algorithm or an Euclidean distance algorithm.

5. The APC system as claimed in claim 1, wherein said R2R controller is a moving-average (MA) controller, an exponentially-weighted-moving-average (EWMA) controller, a double-EWMA controller, or a proportional-integral-derivative (PID) controller.

6. An APC method, comprising:

obtaining a plurality of sets of historical process data used by a process tool for processing a plurality of historical workpieces;

obtaining a plurality of historical measurement data of the historical workpieces measured by a metrology tool, wherein said historical measurement values are the measurement values of said historical workpieces which are manufactured in accordance with said sets of historical process data, respectively;

establishing a conjecture model in accordance with a conjecture algorithm by using said sets of historical process data and said historical measurement values corresponding thereto, and a reference model in accordance with a reference algorithm by using said sets of historical process data and said historical measurement values corresponding thereto, wherein said conjecture algorithm is different from said reference algorithm;

enabling a run-to-run (R2R) controller to control said process tool to perform process runs in accordance with the following relationships:

u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z)

u _k+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,ŷ _k)

G _2,i =f(RI _k)×G _1,i;

where G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning said R2R controller, if RI_k<RI_T;

f(RI _k)=RI_k, if RI_k≧RI_Tand k≦C;

f(RI_k)=1−RI _k, if RI_k≧RI_Tand k>C;

wherein y_zrepresents an actual measurement value of a sampling workpiece which has been processed in the z^thprocess run by said process tool and measured by said metrology tool;

u_z+1represents the control action of the (z+1)^hprocess run when y_zis adopted;

ŷ_krepresents a virtual metrology value of a workpiece which has been processed in the k^thprocess run by said process tool, and is generated by inputting a set of process data of said workpiece into said conjecture model;

RI_krepresents a reliance index (RI) of the kth process run, wherein said reliance index corresponding to the kth process run is generated by calculating an overlap area between the statistical distribution of said virtual metrology value of said workpiece and the statistical distribution of a reference prediction value of said workpiece, wherein said reference prediction value of the kth process run is generated by inputting said set of process data of said workpiece into said reference model, wherein said reliance index is higher when said overlap area is larger, representing that said reliance level of said virtual metrology value corresponding to said reliance index is higher:

C stands for a predetermined number of process runs.

7. The APC method as claimed in claim 6, wherein said conjecture algorithm and said reference algorithm are respectively selected from the group consisting of a multi-regression (MR) algorithm, a support-vector-regression (SVR) algorithm, a neural-networks (NN) algorithm, a partial-least-squares regression (PLSR) algorithm, and a Gaussian-process-regression (GPR) algorithm.

8. The APC method as claimed in claim 6, further comprising:

establishing a statistical distance model is built in accordance with a statistical distance algorithm by using said sets of historical process data; and

enabling said run-to-run (R2R) controller to control said process tool to perform process runs in accordance with the following relationship:

G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning said. R2R controller, if GSI_k>GSI_T, where GSI_krepresents a global similarity index (GSI) of the kth process run and is generated by inputting said set of process data of said workpiece into said statistical distance model; GSI_Trepresents the GSI threshold value defined by two to three times of the maximal global similarity indexes of said sets of historical process data.

9. The APC method as claimed in claim 8, wherein said statistical distance algorithm is a Mahalanobis distance algorithm or an Euclidean distance algorithm.

10. The APC method as claimed in claim 6, wherein said R2R controller is a moving-average (MA) controller, an exponentially-weighted-moving-average (EWMA) controller, a double-EWMA controller (d-EWMA), or a proportional-integral-derivative (PID) controller.

11. A computer program product, which, when executed, performs an APC method comprising:

obtaining a plurality of historical measurement data of the historical workpieces measured by a metrology tool, wherein said historidal measurement values are the measurement values of said historical workpieces which are manufactured in accordance with said sets of historical process data, respectively;

enabling a run-to-run (R2R) controller to control said process tool to perform said process runs in accordance with the following relationships:

u _z+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,y _z)

u _k+1 =g(G _1,1 ,G _1,2 , . . . ,G _1,i ,ŷ _k)

G _2,i =f(RI _k)×G _1,i;

f(RI _k)=RI _k, if RI_k≧RI_Tand k≦C;

f(RI _k)=1−RI _k, if RI_k≧RI_Tand k>C;

represents the controller gain used in said R2R controller when y_zis adopted, wherein i represents the number of the controller gains used in said R2R controller;

ŷ_krepresents a virtual metrology value of a workpiece which has been processed in the kth process run by said process tool, and is generated by inputting a set of process data of said workpiece into said conjecture model;

RI_krepresents a reliance index (RI) of the kth process run, wherein said reliance index corresponding to the kth process run is generated by calculating to an overlap area between the statistical distribution of said virtual metrology value of said workpiece and the statistical distribution of a reference prediction value of said workpiece, wherein said reference prediction value of the kth process run is generated by inputting said set of process data of said workpiece into said reference model, wherein said reliance index is higher when said overlap area is larger, representing that said reliance level of said virtual metrology value corresponding to said reliance index is higher;

RI_Trepresents a RI threshold value based on a maximal tolerable error limit defined by the errors of the virtual metrology values obtained from said conjecture model; and

C stands for a predetermined number of process runs.

12. The computer program product as claimed in claim 11, wherein said conjecture algorithm and said reference algorithm are respectively selected from the group consisting of a multi-regression (MR) algorithm, a support-vector-regression (SVR) algorithm, a neural-networks (NN) algorithm, a partial-least-squares regression (PLSR) algorithm, and a Gaussian-process-regression (GPR) algorithm.

13. The computer program product as claimed in claim 11, said APC method further comprising:

G_2,i=0 or ŷ_k−1but not ŷ_kis adopted for tuning said R2R controller, if GSI_k>GSI_T, where GSI_krepresents a global similarity index (GSI) of the kth process run and is generated by inputting said set of process data of said workpiece into said statistical distance model; GSI_Trepresents a GSI threshold value defined by two to three times of the maximal global similarity indexes of said sets of historical process data.

14. The computer program product as claimed in claim 13, wherein said statistical distance algorithm is a Mahalanobis distance algorithm or an Euclidean distance algorithm.

15. The computer program product as claimed in claim 11, wherein said R2R controller is a moving-average (MA) controller, an exponentially-weighted-moving-average (EWMA) controller, a double-EWMA controller, or a proportional-integral-derivative (PID) controller.